98.

[Cover] [Contents] [Index]

Page 186

Chapter 5
Texture quantisation

Human beings use both spectral and spatial features to interpret visual information. Spectral features describe variations in tone over a greyscale image, whilst spatial features reflect the spatial distribution of these tonal variations, which contain two kinds of spatial relationship. One such relationship is tonal variation focused upon the object of interest, representing the structure of the object, while the other measures the broaderscale relationship between the object being analysed and the remainder of the scene. Generally, the words texture and context are used to represent these two forms of spatial relationship. This chapter focuses on the problem of quantifying texture. The use of context-based methods is described in Chapter 6.

Spectral and textural features are interdependent for, as Haralick et al. (1973) note,‘…texture and tone have an inextricable relationship to one another. Tone and texture are always present in an image, although one property can dominate the other at times’, depending on the ‘fineness’ or ‘roughness’ of the surface of the object, and on the image resolution relative to the surface roughness of the object. If tonal variation inside a limited range is relatively small, spectral information will dominate. For example, at a given resolution, an image may include areas of relatively constant tone, such as water bodies or expanses of concrete or tarmac. Conversely, if tonal variation is large and presents meaningful structures, then texture will be dominant, as in images of rocky areas, settlements and some kinds of cloud.

Texture is an innate property of objects. It contains important information about the structural arrangement of surfaces. The use of texture in addition to spectral features for image classification might be expected to result in some level of accuracy improvement, depending on the spatial resolution of the sensor and the size of the homogeneous area being classified. Where the spatial resolution of the image is fine relative to the scale of tonal variation, texture can be a valuable source of discriminating information. Conversely, where homogeneous regions in the image are small it may prove difficult to estimate texture, for texture is a property of an area

[Cover] [Contents] [Index]